Stocks portfolio optimization based on forecasting through the integration of the MVO method and LSTM model

Abstract

This thesis explores optimal securities portfolio forecasting in the financial market using Long-Short Term Memory (LSTM) models and Mean-Variance Optimization (MVO). LSTM models have proven effective in stock market prediction, while MVO is widely used for portfolio construction. The article "9 of the Best Stocks for a Starter Portfolio" inspired the research. The dataset covers the adjusted closing prices of various stocks from 2010 to April 30, 2023. The LSTM model is trained on 80% of the dataset, with the remaining 20% for testing. The LSTM model consists of two hidden layers with 200 and 100 neurons, a 20% dropout rate, a batch size of 32, and 100 epochs. Daily stock returns are calculated and used for portfolio simulation through MVO. The forecasting performance is evaluated using MAPE, where a value below 2.5% indicates the accuracy of models. The Maximum Sharpe Ratio portfolio is constructed by constructing and analyzing a hypothetical investment portfolio based on different combinations of assets and selecting the one that outperforms the equally-weighted portfolio and the market index (S&P 500) in terms of Sharpe Ratio and Annualized return. Although there is a difference in value, the Maximum Sharpe Ratio portfolio based on predicted data exhibits similar trends to the one based on actual historical data. Asset allocation within the portfolio varies across rebalancing periods, but portfolios constructed with actual and forecasted data are the same. These findings give some insight into combining LSTM models and MVO for optimal portfolio management. The thesis contributes to understanding portfolio optimization by integrating predictive modeling and quantitative techniques.